Cerrado is the second largest biome in Brazil, covering about 2 million km${}^2$. This biome has experienced land use and land cover changes at high rates due to agricultural expansion so that more than $50\%$ of its natural vegetation has already been removed. Therefore, it is crucial to provide technology capable of controlling and monitoring the Cerrado vegetation suppression in order to undertake the environmental conservation policies. Within this context, this work aims to develop a new methodology to detect deforestation in Cerrado through the combination of two Deep Learning (DL) architectures, Long Short-Term Memory (LSTM) and U-Net, and using Landsat and Sentinel image time series. In our proposed method, the LSTM evaluates the time series in relation to the time axis to create a deforestation probability map, which is spatially analyzed by the U-Net algorithm alongside the terrain slope to produce final deforestation maps. The method was applied in two different study areas, which better represent the main deforestation patterns present in Cerrado. The resultant deforestation maps based on cost-free Sentinel-2 images achieved high accuracy metrics, peaking at an overall accuracy of $99.81\%\pm 0.21$ and F1-Score of $0.8795\pm 0.1180$. In addition, the proposed method showed strong potential to automate the PRODES project, which provides the official Cerrado yearly deforestation maps based on visual interpretation.

Near-real-time forest monitoring is a critical component for managing climate risk and resource conservation in the Brazilian Legal Amazon (ALB). The DETER program, managed by the National Institute for Space Research (INPE), has played a pivotal role for over two decades by producing warnings of deforestation and forest degradation to support environmental enforcement by agencies such as IBAMA. However, the effectiveness of these warnings is highly dependent on temporal efficiency — the speed at which a disturbance is detected and published after the activity begins.Recent objective evaluations of DETER’s performance regarding selective logging — a major driver of forest degradation — revealed a significant median delay of approximately 312 days between the start of logging activities and the corresponding warning publication during 2022-2023. This temporal gap highlights the challenge of applying traditional monitoring to complex spatiotemporal datasets, where factors like cloud cover and sensor resolution can hinder early detection.To address these challenges, this research proposes a novel approach within the framework of AI and Machine Learning in Spatiotemporal Contexts. We leverage Foundation Models and Deep Learning architectures designed to process the complex temporal dynamics of tropical forests using Harmonized Landsat-Sentinel (HLS) time series. A key contribution of using foundation models in this pipeline is their ability to learn robust representations from large-scale data, significantly reducing the requirement for vast volumes of manually annotated samples — a known bottleneck for AI-based remote sensing monitoring systems. By applying these models to HLS data, we aim to improve spatiotemporal predictions and the reliability of the modeling pipeline, facilitating the production of more agile and efficient early warnings.This work contributes to the development of the next generation of forest monitoring systems, focusing on interpretability and transferability across the Amazonian landscape. By reducing the detection lag of selective logging, this approach seeks to enhance technological sovereignty in environmental monitoring and provide more effective decision-making support for forest preservation.

Selective logging often marks the beginning of forest degradation, but remote sensing alert systems typically detect it long after it starts. This type of disturbance produces a subtle canopy signal not easily recognised in the early stages, causing alerts in some areas of the Brazilian Legal Amazon to lag behind the first visible signs by several months. This study explores whether a geospatial foundation model can help reduce this time gap. For each documented disturbance site, we compile Harmonized Landsat and Sentinel-2 (HLS) image time series spanning October 2024 to May 2026 into a datacube and use it to pretrain, without labels, a compact spatiotemporal masked autoencoder (ST-MAE). This model learns to embed the dynamics of forest degradation into latent-space representations, the so-called "embeddings". Following this, a lightweight downstream model — trained with few labelled samples — interprets ST-MAE's embeddings to evaluate every new HLS observation for recent logging activity. To effectively translate the evaluation scores into production-ready alerts, we apply a simple change-detection approach: an edge filter for sudden logging-level changes, paired with a persistence filter for sustained detections. We conducted a leave-one-out cross-validation assessment across 20 well-documented logging sites: yielding no false alarms prior to the first disturbance evidence, the change detector confirmed the onset of the 20 logging activities with a median delay of 15 days after the first post-disturbance image. The downstream model's scores remain low during the pre-disturbance period but increase once logging begins, confirming that it responds to logging-related events rather than to other landscape dynamics or calendar artefacts. This label-free pretraining enables effective performance with minimal annotations, providing a practical solution to reduce the temporal gap in monitoring forest degradation.

Biomass burning is a major source of particulate matter and trace gases in South America, yet the extent to which favorable meteorological conditions for fire occurrence translate into atmospheric emissions varies among ecosystems and emission inventories. This study evaluates the ecoregional coupling between the Fire Weather Index (FWI) and biomass-burning total particulate matter (TPM) emissions across 96 Level III ecoregions of South America. Historical daily FWI data for 1981–2024 were combined with TPM emissions from the Global Fire Assimilation System (GFAS), Global Fire Emissions Database (GFED), and Quick Fire Emissions Dataset (QFED) for 2003–2024. Raw monthly FWI and TPM were strongly associated, with median Spearman correlation coefficients (ρ) ranging from 0.73 to 0.76 across inventories. After removal of the mean seasonal cycle, median correlations decreased to 0.42–0.49, indicating that shared seasonality explains part, but not all, of the observed relationship. Positive coupling between FWI anomalies and TPM anomalies remained significant after autocorrelation adjustment and false discovery rate (FDR) correction in 84–88 ecoregions per inventory, while 84 of the 96 ecoregions showed robust agreement across emission-estimation approaches. Monthly mean FWI generally outperformed the frequency of days exceeding the 95th-percentile (P95) fire-weather threshold as an overall predictor of TPM variability; nevertheless, high-emission months contained approximately 5.6–6.6 times more extreme FWI days than other months. Hyper-arid ecoregions exhibited pronounced decoupling despite very high climatological FWI values, consistent with fuel limitation, whereas many tropical, savanna, wetland, and transitional ecoregions showed strong coupling. These results distinguish meteorological fire danger from realized particulate emissions and identify the regions where favorable fire-weather conditions most consistently translate into biomass-burning atmospheric emissions.